ruggsea/bluesky-journalist-classification

Bluesky Journalist Classification Dataset

Dataset Description

This dataset contains Bluesky user profiles for training and evaluating journalist classification models. Created for the CSH Vienna Machine Learning Workshop, it includes comprehensive user data with human-verified labels for binary classification tasks.

Dataset Summary

• Total Examples: 1,189
• Test Split: 229 labeled examples
• Unlabeled Split: 960 unlabeled examples
• Languages: Primarily English
• Task: Binary text classification (journalist vs. non-journalist)
• Collection Date: 2025-07-09

Label Distribution

Test Set

• Journalists: 50 (21.8%)
• Non-Journalists: 179 (78.2%)

Data Collection

The data was collected using the Bluesky AT Protocol API with the following strategy:

1. News Organization Followers (30%): Followers of major news organizations
2. Keyword-Based Search (40%): Users found via journalism-related keywords
3. General Population (30%): Random sampling for balanced representation

Data Sources

• CNN, BBC, New York Times, Reuters, Associated Press followers
• CSH Vienna followers (for general users)
• Keyword searches: "journalist", "reporter", "correspondent", etc.
• General keywords: "developer", "teacher", "artist", "scientist"

Dataset Structure

DatasetDict({
    test: Dataset({
        features: ['user_id', 'handle', 'bio', 'posts_text', 'follower_count', 
                  'following_count', 'post_count', 'is_journalist'],
        num_rows: 229
    }),
    unlabeled: Dataset({
        features: ['user_id', 'handle', 'bio', 'posts_text', 'follower_count',
                  'following_count', 'post_count'],
        num_rows: 960
    })
})

Features

• user_id: Anonymized user identifier
• handle: Bluesky handle (anonymized)
• bio: User profile bio/description
• posts_text: Concatenated recent posts (5-10 posts)
• follower_count: Number of followers
• following_count: Number of accounts following
• post_count: Total number of posts
• is_journalist: Target label (1=journalist, 0=non-journalist)

Usage

from datasets import load_dataset

dataset = load_dataset("ruggsea/bluesky-journalist-classification")

# Use test set for evaluation
test_data = dataset["test"]

# Use unlabeled data for few-shot examples or additional training
unlabeled_data = dataset["unlabeled"]

Classification Approach

This dataset is designed for few-shot learning with large language models:

1. Small Test Set: 229 labeled examples for evaluation
2. Large Unlabeled Set: 960 examples for few-shot prompting
3. Cost-Effective: Use GPT-4o-mini or similar for classification (<$1 for full dataset)

Labeling Guidelines

• Journalist (1): Currently employed by news organization, freelance journalist with published work, independent journalist with regular news content, journalism student actively reporting
• Non-Journalist (0): General public users, academics (unless journalism-focused), business professionals, artists, creators, politicians/activists (unless also journalists)

Ethics & Privacy

• All data from public Bluesky profiles only
• User identities anonymized (handles replaced with generic identifiers)
• No private or protected content included
• Compliant with Bluesky terms of service
• Suitable for academic research and educational purposes

Citation

@dataset{bluesky_journalist_2024,
  title={Bluesky Journalist Classification Dataset},
  author={CSH Vienna},
  year={2024},
  url={https://huggingface.co/datasets/ruggsea/bluesky-journalist-classification}
}

License

• Data: CC BY 4.0 (with attribution)
• Usage: Research and educational purposes